A Fast Data-Driven Method for Submodule Open-Circuit Fault Localization in a Modular Multilevel Converter

The modular multilevel converter (MMC) has become a popular circuit topology in the field of high-voltage direct current (HVdc) technology and has many merits; however, the large numbers of submodules (SMs) are potential failure points. Therefore, an SM open-circuit fault localization (OCFL) strategy can greatly increase the security and reliability of the MMC. The existing OCFL methods commonly have a limited localization speed or are difficult to use in practical engineering cases. A fast and simplified OCFL approach for SMs based on improved unsupervised machine learning is proposed in this article. First, the SM open-circuit fault is analyzed, and a sliding data window and a K-means algorithm are designed. A database is constructed from the continuously sampled SM capacitor voltages. Finally, an improved K-means algorithm is proposed to increase the speed of OCFL. When the faulty data and normal data are clustered, different types of SM open-circuit faults can be quickly and effectively localized. Furthermore, this method does not require extra sensors or precise mathematical models. A three-phase MMC prototype with four SMs per arm is built in the laboratory, and the simulation and experimental results confirm the effectiveness of the proposed method.